End-to-End Molecular Crystal Structure Prediction via Physics-Constrained Retrieval-Augmented GNN-VAE

Crystal Structure Prediction (CSP) remains a formidable challenge in materials science, particularly for organic crystals where the potential energy surface is characterized by polymorphism and complex weak interactions. Traditional methods struggle to balance computational cost with accuracy, often relying on expensive Density Functional Theory (DFT) or inaccurate classical force fields. Furthermore, recent deep learning approaches frequently lack physical constraints, leading to the generation of geometrically valid but thermodynamically unstable structures-a phenomenon we term 'physical hallucination.' To address these limitations, we propose an end-to-end CSP framework integrating a Retrieval-Augmented Generation (RAG) model with a physics-constrained Variational Autoencoder (VAE). We constructed a high-fidelity dataset of 3,737 organic crystal structures, incorporating multidimensional physical data including total energies, stress tensors, and atomic forces calculated via Density Functional Tight Binding (DFTB+). The framework utilizes a dual-encoder architecture (Graph Attention Network and Relational Graph Convolutional Network) to learn a latent representation of crystal stability. Uniquely, we implement a physics-informed loss function that utilizes automatic differentiation to enforce consistency between predicted energies and atomic forces. To enhance generation quality, we introduce RAG to query high-quality structural priors from a knowledge base, guiding Particle Swarm Optimization (PSO) in the latent space. Experimental validation on benzoic acid and anhydrous β-caffeine demonstrates that the model effectively generates thermodynamically stable structures with low Root Mean Square Deviation (RMSD) from experimental benchmarks. This approach offers a robust tool for accelerating organic material discovery by bridging the gap between data-driven generation and physical viability.